Active MIT C# .NET Avalonia UI ONNX Runtime ASP.NET Core Python scikit-learn

ONNX Studio

Cross-platform desktop app to load, inspect, run and serve ONNX and scikit-learn models

GitHub View project →
— Stars
— Forks
— Open issues
— Last push

Features

  • Load ONNX models via file picker, drag-and-drop or CLI, with validation and actionable error messages
  • Inspect the computation graph: category colors, live search and filter, node details, statistics, initializer list
  • Dynamic input forms generated from the model schema (numbers, vectors, images, text)
  • Real local inference with ONNX Runtime, input validation, LRU session cache and typed outputs
  • Embedded REST API (Kestrel): /models, /schema, /predict, /health with CORS and JSON schema/cURL previews
  • API sandbox to send real HTTP requests against the embedded server, with history and re-run
  • Open scikit-learn joblib/pickle models, inspect pipelines and parameters, run inference, and convert to ONNX with skl2onnx
  • VS Code style workbench with Dark+/Light+ themes and keyboard shortcuts

ONNX Studio is a cross-platform desktop application for data scientists, built as a modular monolith on Avalonia, ONNX Runtime and ASP.NET Core. It loads, inspects, runs and serves ONNX models — and scikit-learn models too.

Architecture

The solution is a modular monolith: modules communicate through direct .NET method calls wired by dependency injection, in a single process.

  • ONNXStudio.Core — domain and services, no UI, no HTTP: model loading, registry, inference, LRU session management, form generation, graph analysis, and the Python worker for joblib/pickle models and skl2onnx conversion.
  • ONNXStudio.Api — ASP.NET Core Minimal APIs exposing the loaded models over REST via an embedded Kestrel server.
  • ONNXStudioUI — the Avalonia executable hosting the workbench UI and the API.

Key features

Open a model with Ctrl+O, drag-and-drop, or ONNXStudioUI --model path.onnx. The Inspector shows the computation graph with category colors, live search, node details (attributes, inputs/outputs, dependencies) and statistics. Input forms are generated from the model schema, and inference runs through ONNX Runtime with a typed results view.

The embedded REST API exposes /models, /models/{id}/schema and /models/{id}/predict, with a sandbox to test real HTTP requests and copy cURL previews.

Scikit-learn models saved with joblib or pickle open like any other model: inspect pipeline steps, hyper-parameters and learned attributes, run predict/predict_proba/transform, and optionally convert to ONNX with skl2onnx. Python runs in a separate process, and the runtime can be installed standalone by the app, picked from the system, or pointed to manually.

Build and run

dotnet build ONNXStudio.slnx
dotnet run --project ONNXStudioUI

Pushing a version tag builds tested packages for Windows x64, Linux x64 and macOS Apple Silicon, published as GitHub Releases with installers and SHA-256 checksums.